UHIF: a unified framework via heterogeneous information fusion for flat, nested, and discontinuous NER

Named Entity Recognition (NER), prevalent in real-world applications, has evolved from flat to complex nested and discontinuous structures. However, existing sequence labeling and span-based methods struggle to unify these varied types efficiently and to effectively integrate complementary semantic context and boundary cues. To address these limitations, we propose UHIF, a U nified framework via H eterogeneous I nformation F usion for flat, nested, and discontinuous NER. It leverages a token-pair feature matrix by systematically distinguishing its upper-triangular region for ‘NEXT’ (token-to-immediate-next) labels and its lower-triangular region for ‘SPAN-*’ (head-tail) labels. This representation allows a path-search decoding algorithm to naturally identify flat, nested, and gapped entities simultaneously. To construct this matrix, we employ two specialized modules to capture contextualized semantic information and explicit boundary information, respectively, and leverage the designed Heterogeneous Channel Attention Module (HCAM) to intelligently refine the mixed features—achieving a dynamic balance between semantic and boundary information by modeling cross-channel interactions. Extensive experiments on four diverse datasets (GENIA, CADEC, CoNLL2003, and Resume) demonstrate that UHIF achieves superior performance, obtaining F1 scores of 81.80%, 74.86%, 93.45%, and 96.86%, respectively.